Government AI Records: Build a FOIA-Ready Evidence Map
Government AI teams need a searchable evidence map for authenticated LLM traffic. The map should connect each routed request to agency custody, a records schedule, caller identity, model destination, policy state, and retained content without assigning FOIA exemptions to a gateway.

A caseworker sends benefits notes to an approved LLM through an agency application. Six months later, a FOIA request seeks records about the use of AI in that program. The records team needs to find the source file and the model-bound request. It also needs the response and the policy decision without treating every line of gateway telemetry as automatically releasable. That search succeeds only when the agency designed a custody and retention map before the request arrived.
I think a government AI deployment without that map is unfinished, even when its security controls pass review.
TL;DR
- A FOIA-ready evidence map connects routed LLM events to agency systems and custodians. It also identifies schedules and searchable identifiers.
- Federal law covers recorded information in digital form, but agency officials determine record status and retention. They also determine responsiveness and release.
- DeepInspect can record authenticated HTTP model traffic routed through it. Agency officials decide exemptions, and separate controls cover bypass routes.
Federal records and FOIA answer different questions
The Federal Records Act definition at 44 U.S.C. § 3301 includes recorded information made or received by a federal agency under federal law or in connection with public business when it is preserved or appropriate for preservation as evidence of agency activity or for its informational value. The definition expressly includes information created, manipulated, communicated, or stored in digital form.
FOIA asks a related operational question about records the agency has created or obtained and controls when a request arrives. The Justice Department's Procedural Requirements chapter explains the Supreme Court's two-part agency-record test: creation or acquisition by the agency, plus agency control at the time of the request.
Those rules resist a universal label. A prompt and response may have different owners and schedules from a routing decision and security event. Records officers and counsel make those determinations under agency policy and applicable law. An AI gateway supplies evidence about a bounded traffic path. The agency's designated officials own the records analysis.
FOIA search begins with systems and custody
The FOIA search provisions in 5 U.S.C. § 552 define a search as reviewing agency records manually or by automated means to locate material responsive to a request. The statute also directs agencies to make reasonable efforts to search electronic records. DOJ's guidance describes the adequacy standard in terms of methods reasonably calculated to uncover relevant documents.
A useful evidence map therefore starts with locations and control. For each government AI route, name the source application and business owner. Identify the records custodian and LLM destination. Add the audit store and content repository. Attach the applicable records schedule or unresolved schedule review. Then identify the fields a FOIA professional can query.
Picture the search conference: a yellow legal pad sits beside a laptop while the FOIA officer asks which system holds the actual prompt. A dashboard total leaves that question open, so the map should point to the repository and explain who controls it. It should also state what was intentionally excluded from retention.
This is narrower than the program-level evidence in public-sector AI audit trails. It is a retrieval map for records created or obtained along the routed LLM path.
The evidence map needs stable joins
A model interaction can produce several artifacts. The source case system may hold the document sent for summarization. The calling application may retain a workflow record. The gateway may store a policy decision. A provider account may expose its own activity history. Keeping all content in every system creates unnecessary copies and wider access. Keeping only a model name and timestamp makes later retrieval guesswork.
Use stable, protected joins instead. The gateway record can carry a request identifier and authenticated subject. It can also carry the source application and approved use case. Destination endpoint, model identifier, policy version, disposition, and precise event time complete the route record. A protected case or transaction reference can connect the event to the source repository. Content classification and a digest can support correlation. The agency's retention choice determines whether the underlying text remains available.
The AI decision records pattern provides a reusable event shape. A FOIA-ready map adds custody and repository around that event. It also adds schedule, access, and search instructions. It should also document clock sources and identifier formats so a records team can reconcile systems without relying on one employee's memory.
Retention should follow a documented records decision
Default platform retention is a technical setting, not a records analysis. The records owner should decide which artifacts are records, which repository holds the recordkeeping copy, how long the approved schedule keeps it, and what legal holds suspend disposition. Security telemetry can follow a separate schedule when its purpose and content differ.
Raw prompts and responses deserve explicit treatment. They may contain PII, controlled unclassified information, procurement material, or details about an individual benefit file. Retaining full content can support reconstruction and search, while also creating another sensitive collection. An agency may instead keep the authorized source record in its case system and retain a protected reference with its classification. The gateway can retain the associated policy outcome. That design works only when the source reference remains resolvable for the approved retention period.
The broader government AI architecture places request records beside FedRAMP boundaries and identity systems. It also places them beside agency schedules. The FOIA evidence map should name excluded routes too, including local models and direct provider access. It should also name vendor-native inference that never crosses the managed HTTP path.
Release review remains a human legal process
FOIA makes records available subject to its procedures and nine exemptions. Section 552 also requires agencies to consider partial disclosure and take reasonable steps to segregate and release nonexempt information. DOJ maintains a rolling guide to FOIA procedure, exemptions, and litigation because those decisions depend on legal authority and current case law.
The evidence map should preserve facts needed for that review without pre-deciding the outcome. It can identify likely sensitive data classes and the system owner. It can also identify the provider and contract reference. The related source record should remain linked. The FOIA office still determines responsiveness, applies exemptions, performs foreseeable-harm analysis where required, and decides segregation. Counsel handles privilege questions. Records staff handle schedule and hold questions.
A policy engine that stamps "FOIA exempt" on an event would collapse several legal decisions into a runtime label, and I would reject that design. Useful automation covers retrieval and deduplication. It also covers access control and lineage. Release authority stays with designated agency officials.
Test the map before a request arrives
Choose one approved government LLM route and run a synthetic interaction. Use a marker such as FOIA-MAP-214 in a nonproduction case file. Preserve the source record under the test schedule, send it through an authenticated application session, and confirm that the gateway event carries the expected route and identity. Confirm the policy version, classification, and protected case reference too.
Then hand only the evidence-map entry to a records analyst. The analyst should locate the source record and the associated model event. The analyst should then identify the custodian and find the governing schedule. Repeat the test with a blocked destination. The denial should be searchable even though no provider response exists.
Run a bypass test separately. A direct call outside the managed route should produce no gateway record and should appear as an explicit coverage gap, not a silent success. This same discipline appears in the government agency AI gateway, where route ownership and off-boundary systems remain named.
DeepInspect
DeepInspect is a stateless proxy for authenticated HTTP traffic deliberately routed between agency users or agents and LLM endpoints. It evaluates application-supplied identity and request classification before forwarding. It also evaluates the approved destination and active policy. Each permit, redaction, reroute, or block can produce a signed, tamper-evident decision record outside the calling application's write path.
For a FOIA-ready evidence map, that record supplies the routed event identifier and caller context. It also supplies the destination and policy version. The disposition and time complete the record. The agency can attach protected source references and retention metadata, then join the event to its recordkeeping system. DeepInspect covers only traffic that traverses this authenticated HTTP boundary. Agency systems own source records and schedules. They also own legal holds and FOIA searches. Designated agency officials retain responsibility for exemption analysis, segregation, and release decisions.
Book a technical deep dive at deepinspect.ai.
Frequently asked questions
- Does FOIA require an agency to create a new AI audit log?
FOIA governs access to agency records and the search for responsive records. Separate records-management authorities govern preservation duties, including which records the agency must maintain. The practical design goal is to identify the records an agency creates or obtains through its AI operations. The agency should preserve recordkeeping copies under approved schedules and make those systems searchable. Counsel and records officials should approve that design for the agency and program.
- Is every prompt an agency record?
The answer depends on creation or receipt and agency control. Purpose and the applicable records-management analysis also matter. A prompt used in public business may qualify, but a universal statement would skip the legal and factual test. The evidence map records where prompts exist and who controls them. It records which schedule applies after the agency makes that determination.
- Should the gateway retain full prompt and response text?
That is a custody and minimization decision. Full text can improve reconstruction, but it duplicates sensitive content and expands the access surface. A protected source reference with classification and digest may be sufficient when the authoritative content remains retrievable in the case system. The route, policy, and outcome can remain in the gateway record. The agency should document the choice and test the join.
- Can DeepInspect decide which FOIA exemption applies?
FOIA officials and counsel determine responsiveness and exemptions. They also determine foreseeable harm, segregation, and release. DeepInspect can enforce route and data policies on authenticated LLM HTTP traffic and preserve the resulting decision record. Its role in that process is supplying traffic evidence.